Molecular Generation

Molecular generation produces novel molecular structures—as strings, graphs, or 3D coordinates—often under constraints such as validity, synthesizability, or target properties.

Conditional molecular generation can be written as:

$$m\sim p_{\theta }(m\mid c)$$

where $m$ is a molecule and $c$ can represent a target property, scaffold, protein pocket, or design constraint.

Representation Choices

Molecules can be generated in several spaces:

• SMILES or SELFIES strings.
• Molecular graphs.
• Fragments or reaction templates.
• 3D conformers or protein-ligand poses.
• Latent variables decoded into molecules.

Each representation changes the constraint problem. A string model must learn syntax. A graph model must enforce valence and connectivity. A 3D model must respect geometry, chirality, and conformational uncertainty.

Objective View

A property-conditioned generator often seeks samples that balance likelihood and utility:

$$m^{\text{*}}=\arg \underset{m}{\max }\left[\log p_{\theta }(m\mid c)+\lambda U(m)-\gamma C(m)\right]$$

where $U(m)$ is a utility such as predicted activity and $C(m)$ is a penalty for invalidity, synthetic difficulty, toxicity proxy, or violating design constraints.

For structure-based generation, conditioning may include a pocket representation:

$$m\sim p_{\theta }(m\mid P,c)$$

where $P$ is a protein pocket or receptor context. This makes split design and template leakage especially important.

Why It Matters

• A core tool for de novo design in drug discovery and materials.
• Spans autoregressive, VAE, flow, and diffusion approaches.
• Useful output requires validity and property control, not just novelty.
• Optimization against learned predictors can exploit predictor artifacts.

Evaluation

Generated molecules should be evaluated after standardization:

$$m\stackrel{\text{standardize}}{\to }\tilde{m}\stackrel{\text{dedup}}{\to }\text{validity/novelty/property checks}$$

Common metrics include validity, uniqueness, novelty, diversity, property distribution, scaffold distribution, synthesizability proxies, and domain-specific hit-rate. For structure-based generation, pose quality, steric clashes, interaction plausibility, and receptor split policy also matter.

Validity alone is weak. A generator that produces many valid molecules can still fail by memorizing training scaffolds, exploiting a biased scoring model, or generating molecules outside the applicability domain of the property predictor.

Checks

• Are generated molecules valid, novel, and synthesizable?
• Does the representation (SMILES, graph, 3D) suit the objective?
• Is property conditioning evaluated against held-out targets?
• Are molecules standardized before duplicate and novelty checks?
• Is novelty computed against the train set, known databases, or both?
• Is the property model calibrated and in-domain for generated samples?
• Are ligand scaffolds, protein families, and templates separated when claiming structure-based generalization?
• Are decoy-like shortcuts or trivial property biases ruled out with cheap baselines?

Related

• SMILES
• Molecular graph
• Molecular standardization
• Molecular fingerprint
• Protein-ligand split
• Template leakage
• Pose quality
• Diffusion model
• Autoregressive model
• VAE
• Flow matching
• Protein design
• Equivariant GNN